Methods for real-time estimation of battery anode potential
By estimating the anode potential of lithium-ion batteries using a machine learning framework and controlling the charging current, the problem of lithium deposition caused by fast charging was solved, resulting in faster charging speeds and longer battery life. This demonstrated high fidelity and robustness under different operating conditions.
Patent Information
- Application Number
- CN202580027069.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-04-15
- Filing Date
- 2025-02-28
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to achieve rapid charging of lithium-ion batteries without compromising battery life, leading to accelerated lithium deposition and impacting battery performance and safety.
By employing a machine learning framework, the anode potential of the battery is estimated using measurable signals (current, voltage, and temperature) from the battery management system. Lithium deposition is prevented by controlling the charging current. Combined with the first and second machine learning systems, health status indicators are monitored and updated in real time, enabling fast charging and long lifespan.
It achieves a 30% increase in charging speed and doubles the lifespan without sacrificing battery life, providing higher charging efficiency and safety.
Smart Images

Figure CN123125501A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to methods and systems for estimating the anode potential of a battery in real time in relation to battery charging events.
[0002] The real-time estimation of the anode potential based on this disclosure can be achieved in any type of application with fast-charging batteries, such as electric vehicles, electric ships, electric trucks, electric buses, electric work vehicles, electric aircraft, etc. Background Technology
[0003] Making lithium-ion batteries charge faster and last longer is a known dilemma in the automotive industry. Overemphasizing charging speed can accelerate side reactions within the battery. One of the most critical issues during charging is on-site lithium deposition, which accelerates battery degradation.
[0004] Therefore, there is a need for an improved method to reduce the risk of accelerated battery degradation associated with fast charging. Summary of the Invention
[0005] The purpose of this disclosure is to provide a method and system that avoids the aforementioned problems. This purpose is achieved, at least in part, by the features of the independent claims. The dependent claims contain further improvements to the method and system.
[0006] According to a first aspect of this disclosure, a method is provided for real-time estimation of the anode potential of a battery in relation to a battery charging event, the method comprising: acquiring at least one quantitative state of health (SoH) indicator of the battery as output from a first machine learning system; and feeding the acquired at least one quantitative SoH indicator together with battery charging data associated with the charging event as input to a second machine learning system, the second machine learning system providing an estimation result of the anode potential of the battery as output.
[0007] Lithium deposition occurs when the anode potential of a battery becomes negative. However, installing a sensor for directly measuring the anode potential of a battery is complex. Therefore, accurate and reliable data-driven models have been developed for estimating the anode potential.
[0008] Specifically, this disclosure provides a machine learning framework that estimates the anode potential of the battery using only measurable signals (current, voltage, and temperature) from the battery management system. This information can then be used to control the charging current to keep the anode potential always positive and to prevent on-site lithium deposition.
[0009] The comprehensive machine-learning (ML) framework consists of a first ML system and a second ML system, developed to estimate the anode potential of a battery. Specifically, the anode potential is only used when the vehicle requires fast charging. The first ML system is used primarily for battery aging state estimation and monitoring based on whether the vehicle is being charged to its destination or at home (lower charging current). Then, using the output from the first ML system along with on-site current, voltage, and temperature measurements, the second ML system was developed, which is capable of estimating the battery's anode potential based on the SoH output of the first ML system.
[0010] Lithium deposition typically occurs only when the battery's anode potential becomes negative or near negative, or at least primarily under these conditions. Therefore, based on the anode potential estimation results described above, the charging current can be controlled to avoid deposition and related degradation, which means achieving both faster charging speeds and longer battery life.
[0011] In summary, a systematic machine learning framework was developed to estimate the anode potential of a battery, thereby achieving the fastest possible charging without significantly sacrificing battery life. Furthermore, this method can be used throughout the battery's lifespan to cover the gradual aging process.
[0012] Other advantages can be achieved by implementing one or more features of the dependent claims.
[0013] In some exemplary embodiments that can be combined with any or more of the above embodiments, the first machine learning system frequently updates at least one quantized SoH metric in relation to a charging event based on battery charging data obtained from the charging event. This ensures that one or more quantified SoH metrics remain updated throughout the battery's lifespan.
[0014] In some exemplary embodiments that can be combined with any or more of the above embodiments, the first machine learning system updates at least one quantized SoH metric primarily based on, or solely based on, battery charging data obtained from charging events characterized by slow charging speeds. This ensures that the quality of the estimated quantized SoH metric is improved.
[0015] In some exemplary embodiments that can be combined with any or more of the above embodiments, the second machine learning system estimates the battery's anode potential primarily based on, or solely based on, battery charging data obtained from charging events characterized by fast charging speeds. Lithium deposition is primarily a concern during fast charging, therefore anode potential estimation may not be performed during slow charging events.
[0016] In some exemplary embodiments that can be combined with any one or more of the above embodiments, at least one SoH metric obtained as output from the first machine learning system includes battery capacity as a first SoH metric and at least a first aging mode as a second SoH metric.
[0017] In some exemplary embodiments that can be combined with any one or more of the above embodiments, at least one SoH metric obtained as output from the first machine learning system additionally includes a second aging mode as a third SoH metric, and more specifically, includes a third aging mode as a fourth SoH metric.
[0018] In some exemplary embodiments that can be combined with any one or more of the above embodiments, the first aging mode may be lithium inventory loss (LLI), loss of active material in the negative electrode (LAM), etc. NE ), and loss of active material in the positive electrode (LAM) PE Any one of them.
[0019] In some exemplary embodiments that can be combined with any one or more of the above embodiments, the first aging mode and the second aging mode are lithium inventory loss (LLI) and active material loss in the negative electrode (LAM). NE ), and loss of active material in the positive electrode (LAM) PE Any one of them.
[0020] In some exemplary embodiments that can be combined with any one or more of the above embodiments, the first aging mode, the second aging mode, and the third aging mode are lithium inventory loss (LLI), active material loss in the negative electrode (LAM), and other related aging modes. NE ), and loss of active material in the positive electrode (LAM) PE ).
[0021] In some exemplary embodiments that can be combined with any or more of the above embodiments, at least one SoH metric obtained as output from the first machine learning system includes battery capacity as the first SoH metric and at least a first aging mode as the second SoH metric, wherein the first machine learning system includes a first machine learning model for estimating battery capacity and a second machine learning model for estimating the first aging mode. Thus, dedicated machine learning models can be used to estimate the various SoH metrics, thereby helping to improve the estimation quality.
[0022] In some exemplary embodiments that can be combined with any or more of the above embodiments, at least one SoH metric obtained as output from the first machine learning system includes battery capacity as a first SoH metric and at least a first aging mode as a second SoH metric, wherein the first machine learning system includes a single machine learning model for estimating both battery capacity and the first aging mode. This provides a less complex system implementation.
[0023] In some exemplary embodiments that can be combined with any one or more of the above embodiments, at least one SoH metric obtained as output from the first machine learning system includes battery capacity as the first SoH metric, a first aging mode as the second SoH metric, a second aging mode as the third SoH metric, and more specifically, a third aging mode as the fourth SoH metric, wherein the first machine learning system includes a first machine learning model for estimating battery capacity and a second machine learning model for estimating all aging modes.
[0024] In some exemplary embodiments that can be combined with any one or more of the above embodiments, at least one SoH metric obtained as output from the first machine learning system includes battery capacity as a first SoH metric, a first aging mode as a second SoH metric, a second aging mode as a third SoH metric, and specifically, a third aging mode as a fourth SoH metric, wherein the first machine learning system includes a single machine learning model for estimating both battery capacity and all aging modes. This provides a less complex system implementation.
[0025] In some exemplary embodiments that can be combined with any or more of the above embodiments, at least one SoH metric obtained as output from the first machine learning system includes battery capacity as a first SoH metric, a first aging mode as a second SoH metric, a second aging mode as a third SoH metric, and specifically, a third aging mode as a fourth SoH metric. The first machine learning system includes independent machine learning models for individually estimating each SoH metric, such as Gaussian process regression (GPR), Bayesian regression (BR), random forest regression (RFR), or a neural network (NN). Thus, dedicated machine learning models can be used to estimate each SoH metric, thereby helping to improve the estimation quality.
[0026] In some exemplary embodiments that can be combined with any one or more of the above embodiments, the second machine learning system is implemented as a long short-term memory (LSTM) network machine learning algorithm, an underestimate-enhanced long short-term memory (UE-LSTM) machine learning algorithm, or a neural network, specifically a gated recurrent unit (GRU) neural network or an artificial neural network (ANN).
[0027] In some exemplary embodiments that can be combined with any or more of the above embodiments, battery charging data includes charging current, charging voltage, and battery charging temperature. This battery charging data can be obtained relatively easily.
[0028] In some exemplary embodiments that can be combined with any or more of the above embodiments, the method further includes acquiring battery charging data from a battery management system (BMS) configured to manage the battery. Thus, the system and method according to this disclosure can be readily implemented in a battery charging system.
[0029] In some exemplary embodiments that can be combined with any or more of the above embodiments, the first machine learning system is trained based on battery charging data obtained from charging cycles characterized by slow charging speeds.
[0030] In some exemplary embodiments that can be combined with any or more of the above embodiments, the second machine learning system is trained based on battery charging data acquired from charging cycles characterized by fast charging speeds. Lithium deposition is primarily a concern during fast charging, so the second machine learning system can be trained and used to estimate the anode potential during fast charging events.
[0031] This disclosure also relates to a method for performing fast charging of a battery while reducing the risk of lithium deposition, wherein the method includes controlling the charging current such that the battery anode potential is above a threshold, wherein the battery anode potential is estimated according to any of the methods described above. Battery lithium deposition is generally closely related to the battery anode potential level.
[0032] This disclosure also relates to a computer program for real-time estimation of the anode potential of a battery in relation to a battery charging event, wherein the computer program includes computer-readable instructions that, when executed by a computer, cause the computer to perform the following actions: acquire at least one quantitative state of health (SoH) indicator of the battery as output from a first machine learning system, and feed the acquired at least one quantitative SoH indicator together with battery charging data associated with the charging event as input to a second machine learning system, the second machine learning system providing an estimation result of the anode potential of the battery as output.
[0033] This disclosure also relates to a system for real-time estimation of the anode potential of a battery in relation to a battery charging event, the system comprising one or more processors configured to: acquire at least one quantified state of health (SoH) metric of the battery as output from a first machine learning system, and feed the acquired at least one quantified SoH metric, together with battery charging data associated with the charging event, as input to a second machine learning system, the second machine learning system providing an estimate of the anode potential of the battery as output.
[0034] Other features and advantages of the invention will become apparent when examined in light of the appended claims and the following description. Those skilled in the art will recognize that different features of this disclosure can be combined to create embodiments other than those expressly described above and below, without departing from the scope of this disclosure.
[0035] In summary, to achieve the transition from fossil fuels to renewable and sustainable transportation, batteries must support fast charging and exhibit longer lifespans—two goals that have historically been at odds. Current charging technologies often sacrifice one property for the other, leading to inconvenience or reduced resource efficiency in battery-powered vehicles. This disclosure addresses this challenge by introducing a novel approach involving real-time monitoring and control of deposition potential throughout the entire lifespan of a lithium-ion battery cell. Experimental results with a three-electrode cell reveal that this approach enables at least a 30% increase in charging speed without sacrificing lifespan; in fact, it doubles the lifespan. To facilitate the commercial application of these findings, a machine learning-based framework for estimating the deposition potential over the entire lifespan is proposed, utilizing readily available battery data from electric vehicles. The resulting model demonstrates high fidelity and robustness under various operating conditions, achieving a mean absolute error of only 3.37 mV. This disclosure outlines a practical method for preventing lithium deposition and achieving the fastest battery charging with a focus on health. Attached Figure Description
[0036] The method and system for estimating anode potential according to this disclosure will now be described in detail with reference to the accompanying drawings, in which...
[0037] Figure 1a This shows the EL-cell three-electrode single-cell holder.
[0038] Figure 1b The diagram shows the internal structure of the cell and the measured signals, including the cell voltage U. cell and anode voltage,
[0039] Figure 1c The current and anode potential curves of the studied monomers at different aging levels under CC-Cη charging are shown.
[0040] Figure 1d This shows the current and anode potential curves of another studied cell under CC-CV charging.
[0041] Figure 1e The aging curves of four cells subjected to two different charging strategies are shown.
[0042] Figure 2 An exemplary embodiment of a machine learning framework for estimating the lifetime deposition potential of a battery is shown.
[0043] Figure 3a The histogram shows the estimated results (for all data samples in the test set) for four SoH metrics for a battery operating under the full SoC window.
[0044] Figure 3b The diagram shows a comparison between the estimated trajectory and the actual value for a randomly selected cell in the estimation results of four SoH metrics for a battery operating within the full SoC window.
[0045] Figure 4a The estimation results for the deposition potential are shown, specifically the estimation error distribution for Model 3.
[0046] Figure 4b The results show the estimation of deposition potential, specifically the cumulative absolute error of Model 3 compared to the four alternative models, each characterized by different input variables.
[0047] Figure 4c The results of the estimation for the deposition potential are shown. Specifically, the deposition potential estimated by Model 3 is compared with the curves experimentally measured for randomly selected monomers over their entire lifetime.
[0048] Figure 4dThe results of the estimation for deposition potential are shown as a violin plot of the estimation errors using the proposed UE-LSTM and conventional LSTM on the test dataset, where the horizontal line represents the average error.
[0049] Figure 4e The results of the estimation for deposition potential are shown, specifically the probability distribution of the cumulative estimation error using the UE-LSTM and LSTM models.
[0050] Figure 4f The results show the estimation of deposition potential, specifically the estimation error and R² score for lithium deposition potential when using measurements contaminated with different levels of noise.
[0051] Figure 5 BEVs or HEVs with battery anode estimation systems as described herein are shown. Detailed Implementation
[0052] Various aspects of this disclosure will be described below in conjunction with the accompanying drawings, in order to illustrate and not limit the disclosure, wherein similar reference numerals denote similar elements, and variations of the described aspects are not limited to the embodiments specifically shown, but are other variations that can be applied to the disclosure.
[0053] Lithium-ion (Li-ion) batteries are at the forefront of the transition to sustainable transportation, renowned for their high energy density, cost-effectiveness, and durability, thus establishing them as the primary power source for electric vehicles (EVs). However, as the industry strives to match the convenience of traditional refueling methods, significant challenges have arisen in developing fast-charging technologies that do not compromise battery life.
[0054] Fast charging introduces the following paradox. On the one hand, it requires higher current levels, leading to excessive heat generation. If this heat is not adequately dissipated through advanced thermal management systems, it can significantly accelerate many battery aging mechanisms, such as solid electrolyte interface (SEI) growth. On the other hand, even if the thermal issue is resolved, the increased current triggers complex mechanical and electrochemical reactions within the battery, further exacerbating its degradation.
[0055] Among the many factors that affect battery degradation during fast charging, lithium deposition becomes a critical issue. This phenomenon is characterized by the deposition of metallic lithium on the anode surface, which directly reduces battery capacity and efficiency by hindering the normal intercalation process.
[0056] The pursuit of accurately measuring and controlling deposition potential highlights the complexity of translating theoretical models into actionable strategies. Traditional approaches, ranging from direct measurement to complex estimation algorithms based on equivalent circuits or physics-based battery models, face practical limitations in terms of cost, complexity, and potential interference with battery operation. This predicament underscores the urgent need for innovative approaches to prevent lithium deposition during rapid charging.
[0057] The integration of cloud-based battery management systems (BMS) and the systematic collection of battery operating data have driven advancements in data-driven approaches. These approaches have demonstrated significant proficiency in diagnosing and predicting battery conditions, including state of charge (SoC), state of health (SoH), and remaining useful life (RUL). However, the application of data-driven strategies specifically for estimating deposition potentials to prevent lithium deposition during charging remains in its early stages.
[0058] This disclosure aims to reconcile theoretical insights with practical applications in fast charging of batteries, driven by two main innovations. First, the effectiveness of the deposition potential-controlled charging method is validated using a three-electrode cell, demonstrating significant improvements over conventional methods in both battery life and charging efficiency. Building on this, an innovative ML-based framework for estimating the entire lifespan of battery deposition potential (also known as battery anode potential) is developed. This framework skillfully leverages real-time time-series measurements to learn electrochemical behavior on rapid timescales, while utilizing online-estimated SoH metrics to learn the impact of aging on deposition potential. The current framework is comprehensively validated through extensive battery cycling data under various operating conditions and aging levels, highlighting its robustness and accuracy. These innovations collectively represent significant progress towards health-focused fast-charging solutions for EV batteries.
[0059] Three-electrode single-cell experiment
[0060] In the experimental setup, EL-CELL PAT-Cell was used to test three-electrode monomers, such as... Figure 1a-b is shown. For assembling the three-electrode cell, type 811 nickel manganese cobalt (NMC) oxide was used as the positive electrode, and artificial graphite was used as the negative electrode; both were supplied by Customcells Itzehoe GmbH. All cell assembly operations were performed in an argon-filled glove box. First, the electrode sheets were die-cut to coin size with a diameter of 18 mm. Then, the positive and negative electrodes were sequentially assembled into the PAT-Cell stack. A borosilicate glass fiber separator (260 μm, Whatman FG / A) was used, with a lithium metal ring serving as a reference electrode to measure the electrode potential during battery operation. The electrolyte used was a solution of 90 μL of lithium hexafluorophosphate in a mixture of ethylene carbonate and dimethyl carbonate (LiPF6 LP50). The nominal capacity of the three-electrode cell is 5 mAh, which will be used as the basis for defining the C-rate in subsequent testing instructions.
[0061] The experimental procedure was conducted in an ESPEC LHU-124 temperature-controlled chamber using a PAT-Stand-1 cell holder and a NEWARE coin cell cycler. The chamber was maintained at a constant temperature of 25°C. After cell assembly, formation cycling was performed following the guidelines provided by Customcells. The formation cycling steps are listed in Table 1. Step 1 was repeated twice, followed by Step 2, which was then repeated twice. To assess the aging condition of the cell, a reference performance test (RPT) was scheduled every 25 normal cycles. During these tests, Step 3 was performed twice, and the average of the two discharge capacities was defined as the reference capacity of the cell.
[0062] Table 1: CC-CV charging and CC discharging steps used in the designed three-electrode single-cell experiment.
[0063] In this disclosure, a total of four monomers were assembled and used in experimental activities. Two of these monomers served as a control group and were charged under constant current-constant voltage (CC-CV) conditions according to step 4 in Table 1. The charging process for these monomers was terminated when the current reached 0.1C or the cumulative charge capacity reached 4mAh (equivalent to 80% of the nominal capacity) (see [reference]). Figure 1d The remaining two cells operate under a lithium-free deposition-charging strategy, where the charging current is controlled to maintain a non-negative voltage between the deposition potential (η) and the reference electrode (Li / Li+), thereby ensuring U NE≥0. This pathway was designated as step 5 and referred to as CC-Cη charging, which constituted the experimental group for more in-depth comparative analysis (e.g. Figure 1c (As shown). Here, compared to CC-CV, CC-Cη charging uses a higher current value in the CC phase, thus ensuring that the two strategies have comparable charging times. This arrangement facilitates a direct comparison of their effects on battery degradation. For all defined tests, data on current, voltage, temperature, and deposition potential were measured and collected at a sampling frequency of 1 Hz.
[0064] Synthetic data generation
[0065] To cover a wide range of operating conditions and different aging trajectories that batteries may encounter in practical applications, synthetic data was generated by simulating a single-particle model (SPMe) that includes electrolyte dynamics and combining it with several aging mechanisms. To date, no universally accepted model can encompass all aging mechanisms that may be triggered in a typical Li-ion battery. In this regard, three main aging mechanisms were considered in this simulation: SEI layer growth in the anode and loss of active material in both the anode and cathode. For a more detailed description of the SPMe aging model used, readers can refer to the following literature. JS Edge, S. O'Kane, R. Prosser, ND Kirkaldy, AN Patel, A. Hales, A.Ghosh, W. Ai, J. Chen, J. Yang, et al., Lithium ion battery degradation: what you need to know, Phys. Chem. Chem. Phys. 23 (14) (2021) 8200-8221; SJMoura, FB Argomedo, R. Klein, A. Mirtabatabaei, M. Krstic, Battery stateestimation for a single particle model with electrolyte dynamics, IEEE Trans.Control Syst. Technol. 25 (2) (2016) 453-468; SE O'Kane, W. Ai, G.Madabattula, D. Alonso-Alvarez, R. Timms, V. Sulzer, JS Edge, B. Wu, GJOffer, M. Marinescu, Lithium-ion battery degradation: how to model it, Phys.Chem. Chem. Phys. 24 (13) (2022) 7909-7922.
[0066] It is worth mentioning that, in addition to these three aging mechanisms, lithium deposition is also one of the main aging mechanisms during the use of Li-ion batteries, especially under conditions of higher charging current, lower temperature, higher SoC level, and in aged cells.
[0067] Model-based simulations offer exceptional efficiency in terms of time and energy consumption, and exhibit significant flexibility in construction. This facilitates the development of advanced health management strategies for next-generation BMS, demonstrating a significant advantage over conventional experimental methods.
[0068] By sequentially varying the aging-related parameters, the aforementioned battery model underwent extensive simulations on the PyBaMM platform under a wide range of operating conditions (see the section "Physical Parameters and Simulation Conditions" for details). The simulations produced a large synthetic dataset consisting of two parts. One part of the dataset includes time-series data of current, voltage, temperature, and deposition potential recorded at a sampling rate of 1 Hz. The other part contains four SoH metrics recorded at the end of each RPT cycle: battery capacity, lithium inventory loss (LLI), and loss of active material in the negative electrode (LAM). NE ) and loss of active material in the positive electrode (LAM) PE With the end of life defined as 80% of the nominal capacity, it can be seen that the battery lifespan varies from 20 to 2000 cycles depending on the parameter settings. More details regarding the battery model used and its implementation are presented in the chapter "Electrochemical Battery Modeling and Simulation".
[0069] Framework for lifetime estimation of deposition potential
[0070] According to a first embodiment of this disclosure, an innovative framework designed for estimating the entire lifetime of deposition potential is introduced, such as... Figure 2 As shown, the deposition potential estimation over the entire battery lifespan is based on two key tasks performed through the ensemble of three machine learning (ML) models. Specifically, in this exemplary embodiment, the framework employs Model 1 for estimating battery capacity, Model 2 for quantifying three aging modes on a slow timescale, and Model 3 for estimating deposition potential on a fast timescale. This structured framework leverages the strengths of each model to provide an accurate, robust, and real-time estimate of the deposition potential over the battery's lifetime.
[0071] For Model 1, the inputs (also referred to as features) can be manually constructed from time-series and histogram-based raw data, such as the slope of the voltage curve and the cumulative energy throughput. The output of this model is the battery capacity.
[0072] For Model 2, the inputs include time-series current, voltage, and temperature measurements, while the outputs are three aging modes (i.e., LLI, LAM). PE and LAM NEConsidering the nature of the slow aging dynamics, Model 1 and Model 2 were developed based on data collected during the RPT test, as specified in the "Synthetic Data Generation" section, with the RPT test performed every 25 normal cycles. For real-world vehicle applications, data samples for training these two ML models can be collected during destination charging, where the charging current is low (typically below 1 / 8C). This facilitates favorable conditions, such as a wide SoC range, relatively stable temperature, and a gentle lithium diffusion process, to extract high-quality data samples. To label the four SoH metrics as model outputs, the battery capacity can be calibrated according to step 3 of the "Three-Electrode Single Cell Experiment" section, and the three aging modes are calculated from the open-circuit voltage (OCV).
[0073] Model 3, with deposited potential as its sole output, has two inputs. The first input includes time-series current, voltage, and temperature measurements, from which rapid battery dynamics can be learned. Since the ultimate goal is to optimize battery charging performance in real time, input data can be extracted from the corresponding fast-charging steps. The second input consists of four outputs from Models 1 and 2, used to inform Model 3 of the slow aging characteristics of battery health. Essentially, Model 3 combines time-series and non-time-series features to adaptively estimate the deposited potential over the entire battery lifespan. Furthermore, to facilitate subsequent conservative fast-charging control implementations, an underestimation of the deposited potential is preferred over an overestimation. An underestimation-enhanced long short-term memory (UE-LSTM) is proposed to improve the accuracy of deposited potential estimation while inherently biased towards underestimation. This approach not only ensures prediction accuracy but also strategically aligns with the preference for conservative estimation within the context of fast-charging control mechanisms.
[0074] Model 1 and Model 2 can be considered to form the first machine learning system because they are ML systems that provide quantitative estimates of the SoH metric (both Model 1 and Model 2 provide quantitative estimates of the SoH metric as output). As described below, in some exemplary embodiments, the estimation of the SoH metric can actually be performed by a single ML model or by a system of independent ML models.
[0075] Similarly, Model 3 can be considered as forming a second machine learning system because its task is to provide an estimate of the anode potential (deposition potential) as output, based on the battery fast-charging data as the first input and the quantified estimation result of the SoH index as the second input. In some exemplary embodiments, the estimation of the anode potential can actually be performed by a system of independent ML models, for example, by fusing the outputs of multiple independent ML models.
[0076] In the proposed framework, the first and second machine learning systems operate in parallel. At lower current rates, such as those encountered during destination charging, the risk of lithium deposition decreases, thus reducing the need for real-time lithium deposition potential estimation. In this case, it may not be necessary to enable the second machine learning system (Model 3). Conversely, during fast charging, accurate estimation of the deposition potential becomes crucial for preventing lithium deposition, highlighting the key role of the second machine learning system (Model 3). Notably, the framework can be seamlessly integrated with existing vehicle BMSs using readily available measurements. This integration enables the on-site quantification of all these SoH metrics and real-time estimation of the lithium deposition potential.
[0077] Results of three-electrode single-cell test
[0078] By implementing the experiments designed in the chapter "Three-Electrode Monomer Experiment," the aging trajectories of all the studied monomers can be obtained, such as... Figure 1e As shown. This disclosure presents the first experimental results, which directly measure the deposition potential and use it to dynamically control the charging current in a closed-loop manner throughout the battery's entire operating life. It can be seen that at the end of the test, corresponding to 325 normal cycles, both cells in the CC-CV group had exceeded their end-of-life, with their SoH below 80%. Specifically, these cells experienced a sharp capacity loss of 12% of their nominal capacity within the first 50 cycles, attributed to the onset of lithium deposition, as shown by… Figure 1d This is evidenced by the negative deposition potential shown. In contrast, the two monomers in the CC-Cη group were able to maintain a SoH of over 87% after 325 cycles, indicating a significantly longer lifetime compared to the control. This improvement stems from the CC-Cη charging strategy, which achieves this by continuously monitoring the deposition potential and ensuring its non-negativity.
[0079] In addition to extending battery life, the CC-Cη charging strategy also significantly outperforms CC-CV charging in terms of efficiency, reducing charging time by 30% during the initial cycle and by 40% by the 200th cycle. Figure 1c -d is shown.
[0080] The experimental results above demonstrate that real-time control of the charging process based on the deposition potential (η) can significantly extend battery life while achieving faster charging. These findings not only confirm the feasibility of health-focused fastest charging but also indicate how to achieve it. Given that incorporating a reference electrode into each commercial battery is economically impractical, the experimental results in this laboratory strongly demonstrate the necessity of the lifetime estimation of η as described in the section "Framework for Lifetime Estimation of Deposition Potential".
[0081] Typically, estimations of a given charging strategy related to battery degradation and lifespan rely heavily on extensive experimentation involving thousands of cycles, especially considering varying operating conditions. According to Figure 1a The results in -e show that the long-term effects of different charging strategies on battery health can be largely revealed through the η curves in the early cycling stages, significantly reducing the need for extremely time-consuming and expensive experiments. This benefit, in turn, will accelerate the improvement and optimization of existing battery materials and designs to achieve faster charging and longer lifespans.
[0082] In summary, this disclosure provides a computer-implemented method for real-time estimation of the anode potential of a lithium-ion battery in relation to battery charging events, particularly fast charging events. The method includes: acquiring at least one quantitative state of health (SoH) indicator of the battery as output from a first machine learning system; and feeding the acquired at least one quantitative SoH indicator, along with battery charging data associated with the charging event, as input to a second machine learning system, which provides an estimate of the anode potential of the battery as output based on the input.
[0083] Each of the first and second machine learning systems can be an independent machine learning system. Furthermore, battery charging data associated with charging events can be extracted from ongoing or planned charging events.
[0084] The term "in connection with" here means that the calculations used to determine the estimated anode potential of a lithium-ion battery can be performed, for example, shortly before and / or during the actual battery charging event. The estimation can also be repeated during the charging event in response to changes in charging current, etc.
[0085] To ensure that the quantified SoH metric remains updated during battery use, a first machine learning system is configured to frequently update the at least one quantified SoH metric based on battery charging data obtained from the charging events.
[0086] Such updates are typically performed in relation to charging events. For example, the quantified SoH metric can be updated in relation to individual charging events, to each nth charging event (where n is in the range of 2 to 30), or at least relatively frequently. Importantly, updates are performed cyclically throughout the battery's lifespan to keep the quantified SoH metric up-to-date.
[0087] Battery charging data can be obtained from synthetic battery counts, computer-simulated battery data, or other types of offline battery data. Alternatively, or in combination with this, battery charging data can be obtained from real-time (online) charging events of the battery.
[0088] Specifically, the first machine learning system updates the at least one quantified SoH metric primarily based on, or solely based on, battery charging data acquired from charging events characterized by slow charging speeds. The term "slow charging speed" in this paper refers to lower charging power over a longer period, such as destination charging or similar "lower charging rate" events. Destination charging typically involves slow charging during the night or over at least several hours. This contrasts with "fast charging speed," which typically refers to higher charging power over a shorter period, such as charging at a Supercharger station. Lower charging power means, for example, up to 15kW or up to 25kW, while higher charging power means, for example, 15kW and higher, specifically 25kW and higher.
[0089] As mentioned above, deposition typically occurs only during high-power charging events. Therefore, the second machine learning system estimates the battery's anode potential primarily based on, or solely based on, battery charging data obtained from charging events characterized by fast charging speeds. During slow charging events, it is generally not necessary to adjust the charging current to avoid lithium deposition, and therefore, there is no need to calculate the estimated anode potential.
[0090] As described above, at least one SoH metric obtained as output from the first machine learning system may include, for example, battery capacity as a first SoH metric and at least a first aging mode as a second SoH metric.
[0091] Furthermore, at least one SoH metric obtained as output from the first machine learning system may additionally include a second aging mode as a third SoH metric, and specifically, also include a third aging mode as a fourth SoH metric.
[0092] In some exemplary embodiments, the first aging mode may be lithium inventory loss (LLI) or loss of active material in the negative electrode (LAM). NE ), and loss of active material in the positive electrode (LAM) PE Any one of them.
[0093] In some exemplary embodiments, the first aging mode and the second aging mode are lithium inventory loss (LLI) and active material loss in the negative electrode (LAM). NE ), and loss of active material in the positive electrode (LAM) PE Any one of them.
[0094] In some exemplary embodiments, the first aging mode, the second aging mode, and the third aging mode are lithium inventory loss (LLI), loss of active material in the negative electrode (LAM), and loss of active material in the negative electrode. NE ), and loss of active material in the positive electrode (LAM)PE ).
[0095] As described above, at least one SoH indicator obtained as output from the first machine learning system may include battery capacity as the first SoH indicator and at least a first aging mode as the second SoH indicator, wherein the first machine learning system includes a first machine learning model for estimating battery capacity and a second machine learning model for estimating the first aging mode.
[0096] Furthermore, according to another exemplary embodiment, at least one SoH metric obtained as output from the first machine learning system may include battery capacity as the first SoH metric and at least a first aging mode as the second SoH metric, wherein the first machine learning system includes a single machine learning model for estimating both battery capacity and the first aging mode.
[0097] In some exemplary embodiments, at least one SoH metric obtained as output from the first machine learning system includes battery capacity as the first SoH metric, a first aging mode as the second SoH metric, a second aging mode as the third SoH metric, and more specifically, a third aging mode as the fourth SoH metric. The first machine learning system includes a first machine learning model for estimating battery capacity and a second machine learning model for estimating all aging modes.
[0098] According to another exemplary embodiment, at least one SoH metric obtained as output from the first machine learning system includes battery capacity as the first SoH metric, a first aging mode as the second SoH metric, a second aging mode as the third SoH metric, and more specifically, a third aging mode as the fourth SoH metric, wherein the first machine learning system includes a single machine learning model for estimating both battery capacity and all aging modes.
[0099] According to yet another exemplary embodiment, at least one SoH metric obtained as output from the first machine learning system includes battery capacity as the first SoH metric, a first aging mode as the second SoH metric, a second aging mode as the third SoH metric, and more specifically, a third aging mode as the fourth SoH metric. The first machine learning system includes independent machine learning models for individually estimating each SoH metric, such as Gaussian process regression (GPR), Bayesian regression (BR), random forest regression (RFR), or neural network (NN).
[0100] In some exemplary embodiments, the second machine learning system is implemented as a long short-term memory (LSTM) network machine learning algorithm, an underestimate-enhanced long short-term memory (UE-LSTM) machine learning algorithm, or a neural network, specifically a gated recurrent unit (GRU) neural network or an artificial neural network (ANN).
[0101] In some exemplary embodiments, battery charging data includes at least charging current, charging voltage, and battery temperature. Alternatively, other battery parameters may be used, such as the estimated state of charge (SoC) and / or battery internal resistance.
[0102] In some exemplary embodiments, the method further includes obtaining battery charging data from a battery management system (BMS), which is configured to manage the battery.
[0103] The BMS can also be configured, for example, to protect the battery from lower / higher states of charge, monitor the battery's operating status, control the battery environment, and / or perform cell voltage balancing, etc.
[0104] In some exemplary embodiments, the first machine learning system is trained based on battery charging data obtained from charging cycles characterized by slow charging speeds.
[0105] In some exemplary embodiments, the second machine learning system is trained based on battery charging data obtained from charging cycles characterized by fast charging speed.
[0106] For both slow and fast charging speeds, battery charging data can be obtained offline, for example, from synthetic data generation or simulated battery charge / discharge cycles. Alternatively, or in combination with this, battery charging data can be obtained in real-time, online, by sampling relevant battery parameters such as current, voltage, and battery temperature at an appropriate sampling frequency during actual charging events.
[0107] This disclosure also relates to a method for performing fast charging of a battery while reducing the risk of lithium deposition. The method includes controlling the charging current such that the battery anode potential is above a threshold, wherein the battery anode potential is estimated according to the method described above.
[0108] The threshold can be a predetermined fixed value, such as zero or close to zero. For example, the threshold can be in the range of -0.1V to 0.1V. Alternatively, the threshold can be in the range of 0 to 0.3V or higher. Alternatively, the threshold can be a variable value that changes with certain battery-related variables such as battery temperature. In some exemplary embodiments, the charging current is controlled to maintain a non-negative anode potential.
[0109] Estimation results of SoH index
[0110] By implementing Model 1 and Model 2 in the proposed framework within the full SoC working window, the results for estimating the four SoH metrics are shown. Figure 3a The results are shown in the first two rows of Table 2. These models exhibit high estimation fidelity for all SoH indices, achieving a mean absolute error (MAE) of less than 0.4% and a root-mean-square error (RMSE) of less than 0.55%. Furthermore, although LAM... NE The estimation results produced the highest mean error among all metrics, but for 95% of the test dataset, its absolute error remained within 1.07%, highlighting the model's high accuracy and robustness across a wide range of working conditions. LAM NE The estimation accuracy is slightly lower than that of LAM. PE With LAM PE In comparison, LAM NE The accuracy of the estimation results is relatively reduced, which may be attributed to the fact that the open-circuit potential (OCP) of the anode is flatter than that of the cathode, making it more difficult to identify LAM from the current and voltage of individual cells. NE The changes. To further examine this, a single individual was randomly selected. The estimated trajectories for each SoH metric over the entire lifetime are as follows: Figure 3b The results are shown in the figure. It can be seen that the estimation results closely follow the calibration values, with the maximum estimation error falling within the ±2.5% error range.
[0111] In real-world vehicle applications, battery discharge and charge cycles rarely span the full 0-100% SoC window. Instead, batteries typically undergo various partial charge cycles, including the aforementioned destination charging for EVs. This background necessitates a comprehensive estimation of the developed ML models used for quantifying the SoH metric under realistic charging conditions. Therefore, the performance of Model 1 and Model 2 was evaluated in six different use cases, each characterized by a unique SoC window. Table 2 shows the estimation errors for the four SoH metrics in these scenarios. Although the accuracy is slightly reduced compared to the full SoC window scenario, the obtained estimation results maintain high reliability, with the average error for all scenarios and SoH metrics remaining below 1%. These results validate the effectiveness and practical applicability of Model 1 and Model 2 within the proposed estimation framework. Further results regarding the overall estimation error across different use cases are shown below.
[0112] Table 2: Estimation results of battery SoH index under different charging scenarios.
[0113] Lifetime deposition potential estimation
[0114] 1) Noise Specification for Model Training. In the development of Model 3 within the current estimation framework, the inherent uncertainties in the estimation of the SoH metric, as well as the inherent uncertainties in the measurements of current, voltage, and temperature encountered in practical applications, have been considered. This was achieved by augmenting the training dataset with Gaussian white noise, thereby simulating the impact of estimation and measurement inaccuracies on model training. For all four SoH metrics, the standard deviation of the Gaussian noise was defined using the maximum RMSE observed in six partial charging scenarios (highlighted in Table 2) to calibrate the noise, based on the most challenging real-world conditions. Measurement noise for current, voltage, and temperature was introduced based on specifications and empirical insights associated with widely used battery sensor technologies. Specifically, sensor noise with a standard deviation corresponding to level 4 as defined in Table 4 has been simulated.
[0115] 2) Overall performance of Model 3. As depicted in Figure 4, the verification of deposition potential estimation reveals that Model 3 can predict the cell's deposition potential with high accuracy throughout its lifetime, thus limiting the maximum error to below 20mV. This accuracy is... Figure 4a The results are further highlighted in the local magnification section, where the cumulative error histogram demonstrates that 95% of the estimation results maintain an absolute error below 10 mV. Further quantitative analysis, shown in the last row of Table 3, indicates that the model's MAE is 3.37 mV and RMSE is 4.77 mV. These values confirm the high fidelity and robustness of the model in estimating deposition potentials under different operating conditions.
[0116] Figure 4c The detailed case study shown tracks the deposition potential estimates for randomly selected cells throughout their entire lifespan. A magnified view highlights the performance in estimating deposition potentials near 0V, essential for optimizing battery charging performance. This in-depth analysis reveals that the model not only closely follows the actual trajectory of cell deposition potentials but also dynamically adapts to their aging state.
[0117] 3) The impact of different SoH metrics. The four SoH metrics have been integrated into the feature set of the model.
[0118] However, there is no systematic study on the importance of the SoH index alone for deposition potential estimation. To fill this gap, the number of SoH indices used in the model was varied, resulting in four alternative models characterized by different input variables, of which model 3a uses only directly measured current, voltage, and temperature (see Table 3 for details).
[0119] From Table 3 and Figure 4b We observed that model accuracy consistently improved with increasing numbers of SoH metrics added to the feature set, regardless of the size of the tested data sample. Furthermore, by utilizing a comprehensive set of SoH metrics, Model 3 significantly outperformed all controls lacking complete aging information. Compared to Model 3a, Model 3 reduced MAE and RMSE by 22% and 24%, respectively. This enhancement in prediction accuracy confirms the importance of including detailed information related to battery capacity and aging patterns within the estimation framework.
[0120] 4) Impact of different ML models. To estimate the effectiveness of the proposed UE-LSTM in Model 3 development, three alternative ML algorithms were implemented for comparison: gated recurrent unit (GRU) neural network, artificial neural network (ANN), and conventional LSTM (see the section “UE-LSTM for deposit potential estimation”).
[0121] Comparative analysis between LSTM and UE-LSTM Figure 4d As shown in -e. Although the two algorithms have similar overall performance in terms of estimation accuracy, significant differences were observed in the nature of the estimation error. Specifically, 56% of the estimation error generated by UE-LSTM is attributed to underestimation, representing a 7% increase compared to conventional LSTM. This underestimation characteristic of UE-LSTM means that it can provide a greater safety margin in control strategy designs based on the estimated deposition potential.
[0122] Table 3: Estimation results of deposition potential using different feature sets.
[0123] 5) The impact of different measurement noise levels. For practical battery applications, such as EVs, the accuracy of signals measured from the battery often deviates from that of laboratory environments due to the need to balance sensor cost and accuracy. Given these challenges, robustness to sensor noise is crucial for the estimation model developed. For comprehensive estimation, ten Gaussian white noise levels for current, voltage, and temperature measurements were introduced, with standard deviations defined in Table 4.
[0124] Figure 4f The results demonstrate that the estimation error increases almost linearly with increasing sensor noise levels. Notably, even under the worst-case conditions with noise levels far exceeding typical industry standards (10), the developed estimation model maintains a MAE of less than 16 mV. This finding underscores that the proposed full-lifetime estimation framework and its embedded model 3 provide accurate and robust performance even in the presence of significant sensor noise.
[0125] Table 4: Ten levels of Gaussian white noise were introduced in the measurements.
[0126] computational efficiency
[0127] The computational efficiency of the estimation algorithm, which integrates Models 1 through 3, is important for achieving optimal charging control in a vehicle's BMS. The computational load of the algorithm was evaluated using an Nvidia Tesla A100 GPU in a Python environment. To reduce randomness in the estimation, the algorithm was tested more than ten times under identical conditions, and the average computation time is reported in Table 5.
[0128] Despite using a large dataset, training all three developed models took only 1060 seconds, while the testing phase demonstrated significant computational efficiency, completing in 5.27 seconds. The execution time for estimating the deposition potential within a single real-world charging cycle was only 0.16 seconds. Note that these calculations are conservative, as the three models can be executed in parallel to reduce the total required computation time. These results highlight the algorithm's high practicality for real-world battery management applications.
[0129] Table 5: Computation time required by the proposed estimation algorithm.
[0130] in conclusion
[0131] This disclosure bridges a critical gap in EV technology by challenging the conventional compromise between fast charging and lifespan of lithium-ion batteries. Through rigorous experimental validation on three-electrode cells, a deposition potential-controlled charging strategy is established. This strategy not only allows for a minimum 30% increase in charge rate but also significantly extends battery life by preventing lithium deposition, a well-known contributor to premature battery degradation. To apply these findings to commercial EV batteries without a reference electrode, this disclosure introduces a pioneering machine learning-based framework for estimating deposition potentials over battery life. This framework leverages a combination of real-time battery data, estimated SoH (Solar Hysteresis) metrics, and sophisticated modeling techniques to ensure robust and accurate predictions throughout the battery's lifespan under various operating conditions. This research highlights the opportunity to double battery life using ML techniques without compromising charging speed. The findings of this disclosure mark a key advancement in the development of more efficient, durable, and user-friendly EV batteries.
[0132] Implementation
[0133] Battery anode potential estimation methods are typically implemented in a computer or electronic control unit within the vehicle, or remotely in a fixed server solution capable of establishing a data communication channel with the vehicle.
[0134] Such a computer typically includes a computer program for estimating the battery anode potential in real time in relation to a battery charging event. The computer program includes computer-readable instructions that, when executed by the computer, cause the computer to perform the following actions: acquire at least one quantitative state of health (SoH) indicator of the battery as output from a first machine learning system, and feed the acquired at least one quantitative SoH indicator, together with battery charging data associated with the charging event, as input to a second machine learning system, which provides an estimate of the battery anode potential as output.
[0135] This disclosure also relates to a system for real-time estimation of the anode potential of a battery in relation to battery charging events. This system can be implemented in a vehicle, such as the vehicle's BMS, within another electronic controller in the vehicle, or remotely in a cloud solution.
[0136] Figure 5 The vehicle 16 shown is, in particular, a battery electric or hybrid electric vehicle 16, having a front wheel 17, a rear wheel 18, an electric motor 19 with an integrated inverter, a high-voltage battery 20, a BMS 21, and a battery charging port 22.
[0137] A system for estimating the anode potential of a battery in real time includes one or more processors configured to: acquire at least one quantitative state of health (SoH) metric of the battery as output from a first machine learning system; and feed the acquired at least one quantitative SoH metric, together with battery charging data associated with charging events, as input to a second machine learning system, which provides an estimate of the anode potential of the battery as output.
[0138] The one or more processors can be processors of the BMS, processors of another electronic controller of the vehicle, or processors of remote servers of cloud solutions, etc.
[0139] Experimental Procedure
[0140] Electrochemical cell modeling and simulation
[0141] SEI layer growth model
[0142] The liquid electrolyte used in Li-ion monomers becomes unstable during the first charge cycle and reacts with lithium ions and electrons within the electrode to form a passivated solid-electrolyte interface (SEI), resulting in approximately 10% capacity loss. The SEI layer continues to grow during battery use, leading to further capacity loss and increased resistance. Several methods exist for modeling SEI growth. To describe the dynamics of SEI growth, a kinetically constrained model developed in X.-G. Yang, Y. Leng, G. Zhang, S. Ge, C.-Y. Wang, “Modelling of lithium platinginduced aging of lithium-ion batteries: Transition from linear to nonlinear aging,” J. Power Sources 360 (2017) 28-40, is used, where the exchange current density based on SEI side reactions is expressed in the Tafel equation:
[0143]
[0144] In formula (1), i 0,SEI Here, η is the reference exchange current density, α is the transfer coefficient, n is the number of electrons participating in the reaction, F is the Faraday constant, R is the ideal gas constant, and T represents the cell temperature. In formula (2), η SEI It is the overpotential based on the side reaction of SEI, U refIt is the equilibrium anode potential, which is estimated as a function of the solid concentration at the particle surface, η. NE It is an anodic overpotential, and U ref It is the equilibrium potential of the SEI growth reaction.
[0145] LAM model
[0146] Continuous cycling of Li-ion batteries can cause electrode dissolution, particle breakage, and detachment from the binder material, resulting in LAM (Liquid Activated Mass) in both the negative and positive electrodes. The volume fraction of active material in each electrode is determined by... This indicates that the model can be created using the following formula.
[0147]
[0148] in, and These are the minimum and maximum hydraulic stresses, which depend on the particle radius and the average lithium concentration within the particle. Parameters β, m, and σ yield The value is usually obtained by fitting experimental data.
[0149] Physical parameters and simulation conditions
[0150] The SPMe aging battery model used in this simulation was previously parameterized for LG M50T cylindrical battery cells in "Lithium-ion battery degradation: how to model it" by SE O'Kane, W. Ai, G. Madabattula, D. Alonso-Alvarez, R. Timms, V. Sulzer, JS Edge, B. Wu, G. Joffer, M. Marinescu, Phys. Chem. Phys. 24 (13) (2022) 7909-7922. The parameter values from that reference were adopted, and a set of aging-related parameters were gradually changed in different simulations, such as i in Equation (1). 0,SEI And β in formula (3).
[0151] Each cycle of single-cell operation consists of a CC-CV charge followed by a CC discharge. The CC-CV charge has a 4.2V cutoff voltage during the CC phase and a 50mA cutoff current during the CV phase. The CC discharge has a 2.5V cutoff voltage. A one-hour relaxation period is introduced between adjacent charge and discharge phases to simulate battery behavior under calendar aging. To comprehensively cover different operating conditions encountered in real-world scenarios, various charge current rates and ambient temperatures are introduced into the simulation. The simulation is terminated if a single cell has reached the end of its lifespan. An RPT is performed every 50 cycles to obtain pseudo-OCV measurements, where the single cell is charged and discharged at 1 / 10C until the 4.2V cutoff voltage limit is reached.
[0152] Battery Modeling and Implementation Based on ML
[0153] This section introduces methods for developing three ML models within the proposed lifecycle estimation framework.
[0154] KF-based ML model fusion for capacity estimation
[0155] Methods for online capacity estimation have been well established in the literature. A data-driven multi-model fusion method has recently been developed and demonstrated to have high accuracy and practicality under arbitrary vehicle usage conditions, as disclosed in Y. Zhang, T. Wik, J. Bergström, and C. Zou, "State of health estimation for lithium-ion batteries under arbitrary usage using data-driven multi-model fusion," IEEE Trans. Transp. Electrif. (2023). This data-driven multi-model fusion method can be employed in Model 1 to continuously estimate battery capacity over its lifespan. For details of the method, refer to the aforementioned publication; however, for completeness, its key ideas are presented here: First, a set of customized features is carefully selected from time-series current, voltage, and temperature measurements during battery charging. Based on these features, four different ML algorithms are applied to estimate battery capacity, including two probabilistic approaches and two frequencyistic approaches. Second, a single-step capacity prediction model is developed using histogram data related to usage. Finally, the results of all estimation and prediction models are systematically fused using Kalman filtering (KF).
[0156] LSTM for Aging Mode Quantization
[0157] Individual OCV curves or electrochemical impedance spectroscopy (EIS) can be obtained from various RPT tests and used to achieve the goals of Model 2. However, for practical EV applications, obtaining reliable EIS measurements over a wide frequency range or OCV measurements within the full SoC window is challenging. In contrast, extracting pseudo-OCV curves, for example, during destination charging or overnight home charging, is relatively straightforward. Therefore, features from partial charging events within the LSTM model are constructed to quantify three aging modes as the output of Model 2.
[0158] Unlike traditional neural networks (NNs) that take a single data point as input and require manual feature construction, LSTMs take data sequences and perform regression or classification without feature selection and engineering. This characteristic makes LSTMs well-suited for speech recognition, language translation, and robotics control tasks. Considering battery applications, the collected data is often time-series in nature. Therefore, using LSTMs to build estimation models is appropriate. Compared to recurrent neural networks (RNNs), which are also designed to process time-series data, LSTMs effectively avoid gradient vanishing during training.
[0159] A typical LSTM unit consists of two internal cells, namely, the memory cell c. t and hidden cells h t Where t represents the current time step. Memory cells can be considered as states that store information over a long period. Hidden cells, along with a gating mechanism consisting of three main gates (input gate, forget gate, and output gate), control the flow of information within the neural network. The detailed calculation process can be represented by the following formula:
[0160] f t = σ(W f x t + U f h t-1 + b f ), i t = σ(W i x t + U i h t-1 + b i ),
[0161] o t = σ(W o x t + U o h t-1 + bo ), c˜ t = tanh(W c x t + U c h t-1 + b c ),
[0162] c t = f t ⊙ c t-1 + i t ⊙ c˜ t , h t = o t ⊙ tanh(c t ),
[0163] Where, x t It is the input vector, i t f t and o t These are the enable vectors for the input gate, forget gate, and output gate, respectively, c˜ t is the input enable vector for the memory cell. W and U represent the weight matrices of the input state and hidden state, respectively, b is the bias term, and ⊙ represents element-wise product.
[0164] Alternative methods for estimating the SoH index
[0165] As an alternative to the above approach, two other approaches can be used to estimate the four SoH indices. The first is to extend the LSTM model described in the previous paragraph, "LSTM for Aging Pattern Quantization," to include all four SoH indices as output. The second is to extract relevant and appropriate feature sets for each SoH indices and use them to train point-to-point ML models, such as Gaussian process regression (GPR), Bayesian regression (BR), random forest regression (RFR), and neural networks. A systematic comparison of the proposed approaches with their two benchmarks is then performed. In general, the approaches proposed in the previous sections, "KF-based ML model fusion for capacity estimation" and "LSTM for Aging Pattern Quantization," provide the most accurate estimation results for all indices.
[0166] UE-LSTM for Deposition Potential Estimation
[0167] For fast-charging control based on deposited potential, it is crucial to have a small safety margin so that the deposited potential does not enter the negative region. Therefore, it becomes necessary to preferentially employ a model that tends to underestimate rather than overestimate, especially when the deposited potential is close to zero. To achieve this, a new algorithm called UE-LSTM is proposed, which employs a unique loss function J, defined as follows:
[0168]
[0169] Among them, y i and yˆ i These are the actual deposition potential and the estimated deposition potential, respectively. Equation (5) is designed with a weighting factor w1=10 to favor the underestimation scenario. A larger weighting factor w2=50 is introduced in Equation (6) to improve the estimation accuracy in the deposition potential range of [-0.02, 0.02] V.
[0170] Training data labeling
[0171] The capacity of a single battery cell can be measured directly during RPT testing or estimated using an accurate and reliable estimator, such as the estimator developed in the publication "State of healthestimation for lithium-ion batteries under arbitrary usage using data-driven multi-model fusion" by Y. Zhang, T. Wik, J. Bergström, and C. Zou, IEEE Trans. Transp. Electrif. (2023). This publication relates to a data-driven multi-model fusion method for SoH estimation under arbitrary usage steps, where all possible operating conditions are divided into six scenarios, and for each scenario, an appropriate feature set is extracted to indicate SoH. Based on the acquired features, four machine learning algorithms are applied individually to train the SoH estimation model using time-series data. In addition to the estimation at the current time step, an online adaptive model based on histogram data from previous research is adopted to predict the SoH for the next step. Kalman filtering is then applied to systematically fuse the results of all estimation and prediction models.
[0172] As described in Y. Fang, AJ Smith, RW Lindström, G. Lindbergh, and I. Furo, “Quantifying lithium lost to plating and formation of the solid-electrolyteinterphase in graphite and commercial battery components,” Appl. Mater. Today 28 (2022) 101527, there are established methods for experimentally extracting LLI information through advanced laboratory settings. However, it is preferable to continuously calculate these three aging modes throughout the battery life using pseudo-OCV measurement data and diagnostic algorithms such as those described in CR. Birkl, MR. Roberts, E. McTurk, PG. Bruce, and DA. Howey, “Degradation diagnostics for lithium ion cells,” J. Power Sources 341 (2017) 373-386.
[0173] Estimation matrix
[0174] MAE and RMSE were used to quantify the estimated performance for the four SoH metrics under consideration.
[0175] These two estimation matrices are defined as follows:
[0176]
[0177] Where N represents the total number of samples in the test set, y i and yˆ i These represent the true value and the estimated value, respectively, where i is the time step. For deposition potential estimation, in addition to MAE and RMSE, the coefficient of determination is used to estimate the model's fit to the test samples, defined as follows:
[0178]
[0179] Among them, the average estimated value .
[0180] The present disclosure has been illustrated above with reference to specific embodiments. However, other embodiments besides those described above are also possible and are within the scope of this disclosure. Method steps, performed by hardware or software, different from the steps described above, can be provided within the scope of this disclosure.
[0181] The methods disclosed herein can be implemented in a general-purpose computer, processor, or processor core. Suitable processors include, by way of example, general-purpose processors, special-purpose processors, conventional processors, digital signal processors (DSPs), multiple microprocessors, one or more microprocessors associated with a DSP core, controllers, microcontrollers, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), any other type of integrated circuit (IC), and / or state machines.
[0182] The methods or flowcharts provided herein can be implemented as computer programs, software, or firmware incorporated in a computer-readable storage medium for execution by a general-purpose computer or processor. Examples of computer-readable storage media include read-only memory (ROM), random access memory (RAM), registers, cache memory, semiconductor memory devices, magnetic media such as internal hard disks and removable disks, magneto-optical media and optical media such as CD-ROMs, and digital versatile disks (DVDs).
[0183] The processor associated with the system of this disclosure may be, or include, any number of hardware components configured to perform data or signal processing or to execute computer code stored in memory. The system may have associated memory, which may be one or more devices configured to store data and / or computer code for performing or facilitating the various methods described herein. The memory may include volatile or non-volatile memory. The memory may include database components, object code components, script components, or any other type of information structure for supporting the various activities of this specification. According to exemplary embodiments, any distributed or local memory device may be used with the systems and methods of this specification. According to exemplary embodiments, the memory is communicatively connected to the processor (e.g., via circuitry or any other wired, wireless, or network connection) and includes computer code for performing one or more processes described herein.
[0184] It should be understood that the above description is merely exemplary in nature and is not intended to limit the scope of this disclosure, its application, or use. Although specific examples have been described in the specification and illustrated in the drawings, those skilled in the art will understand that various modifications can be made and elements can be replaced with equivalents without departing from the scope of this disclosure as defined in the claims. Furthermore, modifications can be made to adapt particular situations or materials to the teachings of this disclosure without departing from its essential scope.
[0185] Therefore, this disclosure is not intended to be limited to the specific examples illustrated in the drawings and described in the specification as the best mode contemplated for implementing the teachings of this disclosure. Rather, the scope of this disclosure will include any embodiments falling within the scope of the foregoing description and the appended claims. Reference numerals used in the claims should not be considered as limiting the scope of the matters protected by the claims; their sole purpose is to facilitate the understanding of the claims. Figure Labels 1. EL-cell three-electrode single-cell holder 2. Battery cell 3. Synthetic Data Generation 4. RPT process 5. Fast charging process 6. Machine Learning Model for Capacity Estimation 1 7. Machine learning model 2 for quantifying aging patterns 8. Quantitative SoH Indicator 9. Data collection during online deployment, and online measurement of battery parameters. 10. Destination charging (slow charging) 11. Fast charging 12. Machine Learning Model for Lifetime Anode Potential Estimation 3 13. Estimated battery anode potential 14. The First Machine Learning System 15. Second Machine Learning System 16. Battery-electric or hybrid electric vehicles 17. Front wheel 18. Rear wheel 19. Electric motor with integrated inverter 20. High-voltage battery 21.BMS 22. Battery charging port
Claims
1. A method for estimating the anode potential of a battery (20) in real time in relation to a battery charging event, the method comprising: At least one quantitative health status (SoH) indicator of the battery (20) is obtained as output from the first machine learning system (14); as well as The acquired at least one quantized SoH index, together with the battery charging data associated with the charging event, is fed as input to a second machine learning system (15), which provides an estimate of the anode potential of the battery (20) as output.
2. The method according to claim 1, wherein, The first machine learning system (14) frequently updates the at least one quantized SoH metric in relation to the charging event based on battery charging data obtained from the charging event.
3. The method according to any one of the preceding claims, wherein, The first machine learning system (14) updates the at least one quantified SoH index primarily based on or solely based on battery charging data obtained from charging events characterized by slow charging speeds.
4. The method according to any one of the preceding claims, wherein, The second machine learning system (15) estimates the anode potential of the battery (20) primarily based on or solely based on battery charging data obtained from charging events characterized by fast charging speed.
5. The method according to any one of the preceding claims, wherein, The at least one SoH metric obtained as output from the first machine learning system (14) includes battery capacity as a first SoH metric and at least a first aging mode as a second SoH metric.
6. The method according to claim 5, wherein, The at least one SoH metric obtained as output from the first machine learning system (14) additionally includes a second aging mode as a third SoH metric, and specifically also includes a third aging mode as a fourth SoH metric.
7. The method according to any one of claims 5 to 6, wherein, The first aging mode is lithium inventory loss (LLI) and loss of active material in the negative electrode (LAM). NE ), and loss of active material in the positive electrode (LAM) PE either of the following, or The first aging mode and the second aging mode are lithium inventory loss (LLI) and active material loss in the negative electrode (LAM). NE ) and loss of active material in the positive electrode (LAM) PE either of the following, or The first aging mode, the second aging mode, and the third aging mode are lithium inventory loss (LLI) and active material loss in the negative electrode (LAM). NE ), and loss of active material in the positive electrode (LAM) PE ).
8. The method according to any one of claims 5 to 7, wherein, The at least one SoH metric obtained as output from the first machine learning system (14) includes battery capacity as a first SoH metric and at least a first aging mode as a second SoH metric. The first machine learning system includes a first machine learning model for estimating the battery capacity and a second machine learning model for estimating the first aging mode, or The first machine learning system includes a single machine learning model for estimating both the battery capacity and the first aging mode.
9. The method according to any one of claims 6 to 8, wherein, The at least one SoH metric obtained as output from the first machine learning system (14) includes battery capacity as a first SoH metric, a first aging mode as a second SoH metric, a second aging mode as a third SoH metric, and particularly includes a third aging mode as a fourth SoH metric. The first machine learning system includes a first machine learning model for estimating the battery capacity and a second machine learning model for estimating all aging modes, or Wherein, the first machine learning system includes a single machine learning model for estimating both the battery capacity and all aging modes, or The first machine learning system includes independent machine learning models for individually estimating each SoH metric, such as Gaussian process regression (GPR), Bayesian regression (BR), random forest regression (RFR), or neural network (NN).
10. The method according to any one of the preceding claims, wherein, The second machine learning system (15) is implemented as a long short-term memory (LSTM) network machine learning algorithm, an underestimate-enhanced long short-term memory (UE-LSTM) machine learning algorithm, or a neural network, particularly a gated recurrent unit (GRU) neural network or an artificial neural network (ANN).
11. The method according to any one of the preceding claims, wherein, The battery charging data includes charging current, charging voltage, and battery charging temperature.
12. The method according to any one of the preceding claims, wherein, The first machine learning system (14) is trained based on battery charging data obtained from charging cycles characterized by slow charging speeds; and / or The second machine learning system is trained based on battery charging data obtained from charging cycles characterized by fast charging speed.
13. A method for performing fast charging of a battery (20) while reducing the risk of lithium deposition, the method comprising: Controlling the charging current to make the battery anode potential higher than a threshold, wherein the method according to any one of claims 1 to 12 estimates the anode potential of the battery.
14. A computer program for real-time estimation of the anode potential of a battery (20) in relation to a battery charging event, the computer program comprising computer-readable instructions that, when executed by a computer, cause the computer to: At least one quantitative health status (SoH) indicator of the battery (20) is obtained as output from the first machine learning system (14); and The acquired at least one quantized SoH index, together with the battery charging data associated with the charging event, is fed as input to a second machine learning system (15), which provides an estimate of the anode potential of the battery (20) as output.
15. A system for real-time estimation of the anode potential of a battery (20) in relation to a battery charging event, the system comprising one or more processors configured to: At least one quantitative health status (SoH) indicator of the battery (20) is obtained as output from the first machine learning system (14); and The acquired at least one quantized SoH index, together with the battery charging data associated with the charging event, is fed as input to a second machine learning system (15), which provides an estimate of the anode potential of the battery (20) as output.
